O R I G I N A L P A P E R
Scanning reflectance spectroscopy (380–730 nm):
a novel method for quantitative high-resolution climate
reconstructions from minerogenic lake sediments
M. Trachsel•M. Grosjean•D. Schnyder •
C. Kamenik•B. Rein
Received: 17 April 2010 / Accepted: 20 September 2010 / Published online: 7 October 2010 Springer Science+Business Media B.V. 2010
Abstract High-resolution (annual to sub-decadal) quantitative reconstructions of climate variables are needed from a variety of paleoclimate archives across the world to place current climate change in the context of long-term natural climate variability. Rapid, high-resolution, non-destructive scanning techniques are required to produce such high-resolu-tion records from lake sediments. In this study we explored the potential of scanning reflectance spec-troscopy (VIS-RS; 380–730 nm) to produce quanti-tative summer temperature reconstructions from minerogenic sediments of proglacial, annually lam-inated Lake Silvaplana, in the eastern Swiss Alps.
The scanning resolution was 2 mm, which corre-sponded to sediment deposition over 1–2 years. We found correlations up to r = 0.84 (p \ 0.05) for the calibration period 1864–1950, between six reflec-tance-dependent variables and summer (JJAS) tem-perature. These reflectance-dependent variables (e.g. slope of the reflectance 570/630 nm, indicative of illite, biotite and chlorite; minimum reflectance at 690 nm indicative of chlorite) indicate the mineral-ogical composition of the clastic sediments, which is, in turn, related to climate in the catchment of this particular proglacial lake. We used multiple linear regression (MLR) to establish a calibration model that explains 84% of the variance of summer (JJAS) temperature during the calibration period 1864–1950. We then applied the calibration model downcore to develop a quantitative summer temperature recon-struction extending back to AD 1177. This temper-ature reconstruction is in good agreement with two independent temperature reconstructions based on documentary data that extend back to AD 1500 and tree ring data that extend back to AD 1177. This study confirms the great potential of in situ scanning reflectance spectroscopy as a novel non-destructive technique to rapidly acquire high-resolution quanti-tative paleoclimate information from minerogenic lake sediments.
Keywords Sedimentology Limnogeology Climate change Reflectance spectroscopy Clastic sediment Alps Switzerland
M. Trachsel (&) M. Grosjean D. Schnyder C. Kamenik
Department of Geography, University of Bern, Erlachstrasse 9a, 3012 Bern, Switzerland e-mail: trachsel@giub.unibe.ch
M. Trachsel M. Grosjean C. Kamenik Oeschger Centre for Climate Change Research, University of Bern, Za¨hringerstrasse 25, 3012 Bern, Switzerland
B. Rein
Institute for Geosciences, University of Mainz, Becherweg 21, 55099 Mainz, Germany B. Rein
GeoConsult Rein, Gartenstrasse 26-28, 55276 Oppenheim, Germany
Introduction
High-resolution, well calibrated temperature recon-structions are fundamental to put current warming into a longer-term perspective. To improve knowl-edge of past climate and further constrain uncertain-ties in climate reconstructions, new high-quality time series of proxy climate data are required. These proxy series must fulfill several criteria that are fundamental for quantitative reconstructions: (1) the proxy data must be highly correlated with climate variables, (2) data must be collected at high temporal resolution (i.e. annual to multi-year intervals), and (3) the record must be precisely dated.
In the past decade, considerable advances have been made in the use of geochemical proxies in lake sediments for reconstructing past climate. These reconstructions were based on sedimentological or geochemical analysis of thin sections (Francus et al.
2002; Kalugin et al. 2007) or sediment subsamples (McKay et al. 2008; Kaufman et al. 2009). Such analyses were destructive and required time-consum-ing sample preparation. Major developments have also taken place in the exploration of methods of direct measurement on sediment cores using non-destructive scanning techniques such as X-ray fluo-rescence (Zolitschka et al.2001).
Little, however, has been done on scanning reflectance spectroscopy in the visible spectrum (VIS-RS; 380–730 nm), another non-destructive rapid scanning method (Rein and Sirocko 2002). One advantage of this method is that data acquisition is rapid and inexpensive. Several meters of sediment core can be processed at 2-mm measuring intervals in a day (i.e. [1,000 data points can be generated), and the same sediment can be used for further analyses. More difficult are data analysis, interpretation, and transformation of raw reflectance spectroscopy data into quantitative variables that represent specific sediment compounds, and ultimately are informative about climate or environmental change. Reflectance spectroscopy in the visible and infrared range is widely used to detect organic components in lake sediments (Rosen et al.2000; Rein and Sirocko2002; Wolfe et al.2006; Michelutti et al.2009). However, non-destructive scanning reflectance spectroscopy, with direct in situ measurement of split sediment cores is far less common. This method has been used to detect organic and clastic components in marine
sediments off the coast of Peru (Rein and Sirocko
2002; Rein et al.2005). In a recent study, von Gunten et al. (2009) used this scanning technique on biogenic freshwater sediments in eutrophic Laguna Aculeo, central Chile, to determine photopigments and pro-duce high-resolution quantitative reconstructions of austral summer (DJF) temperatures.
In this study we explored the potential for applying this method to minerogenic freshwater sediments in a proglacial lake environment. We chose Lake Silva-plana, eastern Swiss Alps, for this methodological case study because the mineralogical and geochemical composition of the sediment is well known from established analytical techniques (X-ray diffraction, ICP-OES measurements of biogenic Si, grain size distributions and Mass Accumulation Rate (Blass et al.
2007a,band references therein), and the mineralog-ical composition of individual varves has been shown to be controlled by climate (Trachsel et al. 2008). These sediments contain \1% organic material. The rest is composed of mineroclastic fine silt particles that are transported as suspended load from glacial melt-water and rainfall runoff (Blass et al.2007a).
The overall goals of this study were: (1) to systematically test the method and the significance of the data produced, and (2) to describe how raw reflectance measurements can be used to make quantitative climate reconstructions. We use pub-lished algorithms that describe the reflectance char-acteristics of minerals identified in the sediments of Lake Silvaplana using XRD. In addition, we present further algorithms that describe the remaining, as yet unexplained distinct characteristics of the measured sediment reflectance spectra. Each of these algo-rithms is then used to produce a spectrum-derived variable. To gain further insight into the ‘known’ and ‘unknown’ variables extracted, we applied a Principal Component Analysis (PCA) to the variables to detect similarities. Subsequently, we transformed the raw reflectance data (variables) into a climate reconstruc-tion, which involved four steps: (1) testing the effects of climate variables and sediment parameters on the spectrum-derived variables, using redundancy analy-sis RDA, (2) establishing univariate and multivariate calibration models for climate and sediment param-eters, (3) applying the optimal calibration model downcore (reconstruction), and (4) verifying the reconstruction back in time to AD 1177, with fully independent data sets.
Study site
Lake Silvaplana (Lej da Silvaplauna, Fig. 1, 46270N,
9480E) is a postglacial, high-elevation (1,800 m),
2.7-km2 (water volume 127 9 106m3), 77-m deep, dimictic lake of glacio-tectonic origin. The lake is typically ice-covered between January and May. The catchment is 129 km2 and reaches an elevation of 3,441 m. About 6 km2, 5% of the catchment, were glaciated in 1998. The Fedacla River, with average runoff 1.5 m3 s-1, is the only glacial river in the
catchment and, therefore, the principle conveyor of sediments to the lake. The Inn River, connecting Lake Sils with Lake Silvaplana, is larger (discharge 2 m3 s-1), but is almost devoid of sediments. Two small rivers (Vallun; 0.7 m3s-1 and Surlej; 0.3 m3 s-1) are particularly active during spring snowmelt and summer rainstorms.
The specific geological setting of the basin was fundamental to our study. Lake Silvaplana is located on the tectonic Engadine Line, which separates the Lower Austroalpine basement (granites with high
Fig. 1 Overview map of the Lake Silvaplana catchment area, including the coring location, meteorological station, geology and the glacier extent during the Little Ice Age and in 1991 (redrawn from Blass et al.2007a)
amounts of feldspar, Err-Bernina nappe) in the northwest and east from the Penninic basement (orthogneiss and gabbro, with high amounts of mica, Margna nappe and Platta nappe, Fig.1) in the south (Spillmann1993). These spatial differences in geol-ogy are apparent in the mineral spectra of sediments around the shore of the lake (Ohlendorf1998).
The Engadine exhibits climate typical for an inner-Alpine dry valley. Maximum monthly precipitation occurs in August (121 mm), whereas a minimum is observed in February (42 mm). Annual precipitation amounts to 978 mm (1961–1990, SMA2002). Mean monthly temperature ranges from -7.2C (January) to 10.4C (July).
Materials and methods
A gravity core (SVP 06-1) and two piston cores (06-2 and 06-3) were collected from Lake Silvaplana in March 2006. The chronology of the top section (SVP 06-1; AD 1620 onwards) was established by visual correlation of 16 diagnostic sediment strata (nine flood layers for the calibration period 1864–1950; seven flood layers 1620–1864) with the master chronology developed for freeze cores in previous work by varve counting and documented flood layer events (Blass et al. 2007a; Trachsel et al. 2010) and interpolated with a mixed-effect regression model (Heegaard et al.
2005). Maximum error in the chronology was ±7 years. The chronology of cores SVP 06-2 and SVP 06-3 (AD 1177–1620) was established using varve counts and historically documented flood layer events (Caviezel2007), extending the record back to AD 1177 (Trachsel et al.2010).
For numerical calibration analysis we used the data set from the time period 1864–1950 for which monthly temperature and precipitation data are available from the nearby Sils–Maria meteorological station (Begert et al. 2005). After 1950, anthropo-genic eutrophication (Blass et al.2007b; Bigler et al.
2007) strongly affected sediment composition, which in turn affects the reflectance characteristics of the sediment and masks the climate signal (Rein and Sirocko2002; Wolfe et al.2006).
The reflectance data were acquired by high-resolution photospectrometric logging of split cores with a Gretag Spectrolino (GretagMacbeth, Switzer-land). The aperture of the sensor is 2.5 mm and the
spectral coverage ranges from 380 to 730 nm, with a spectral resolution of 3 nm integrated into bands of 10-nm intervals. The photo-spectrometer uses a ceramic plate (BCA–GretagMacbeth) as a white standard. Prior to analysis the split cores were covered with polyethylene transparent foil to mini-mise oxidation and avoid contamination of the sensor head. Wavelength-dependent illumination and trans-parency effects were corrected by dividing each spectrum acquired from the sediment by that of the transparency-covered standard (Rein and Sirocko
2002). The measurement interval on the core was 2 mm.
After data acquisition, diagnostic reflectance char-acteristics, hereafter referred to as ‘variables,’ were extracted from the overall reflectance spectra by applying specific algorithms (Rein and Sirocko2002; Wolfe et al.2006). In Lake Silvaplana the reflectance spectra are combinations of the reflectance spectra of the main minerals present in the sediment, i.e. quartz, mica (muscovite, biotite, illite), chlorite, plagioclase and amphibole (Ohlendorf1998; Trachsel et al.2008). For this study, we used four algorithms that are well established and documented in the literature and are indicative of illite, biotite and chlorite (Rein 2003; USGS 2007), and two algorithms that describe fur-ther characteristics of the spectra but are presently ‘unknown’ (Table1). In general, either slopes of reflectance spectra in defined windows (Rein and Sirocko2002; Wolfe et al.2006), relative absorption band depths (Rein and Sirocko, 2002), or relative absorption band areas (Rein and Sirocko2002; Wolfe et al.2006) were used to describe the spectra (Fig.2b). In general, a relative absorption band depth (RABD, min) is defined as the ratio between the value measured in a reflectance minimum and the value obtained for the same wavelength, assuming a linearly interpolated reflectance continuum (blue line, Fig.2b) between two maxima (Fig.2b). Hence, it is the ratio between the points denoted with a and b in Fig.2b.
Point a can be obtained by distance weighting of the maxima, i.e.
a¼ðx3 x2Þ y1 þ ðx2 x1Þ y3
ðx3 x1Þ ;
or by adding the slope multiplied with the distance, to one maximum:
a¼ y1 þy3 y1
x3 x1 x2 x1ð Þ
¼ y1 ðx3 x1Þ þ ðy3 y1Þ ðx2 x1Þ ðx3 x1Þ
Elementary calculus proves the equality of these two approaches.
Relative absorption band area (RABA, trough) is defined by the area between the measured reflectance and the linearly interpolated reflectance between two maxima described above (red area in Fig. 2b), divided by the overall reflectance. This area is defined by the rectangle constrained by 0 and y1, and x1 and
Table 1 Algorithms for the calculation of the six VIS-RS-derived variables
Algorithm
R590dR690 R590/R690 Biotite, Illite, Chlorite
(USGS2007)
R570dR630 R570/R630 Biotite, Illite, Chlorite
(Rein2003) Min690 [(4*R590? 10*R730)/14]/R690 Chlorite (USGS2007)
Trough590–730 (141*R730? [141*(R590- R730)/2] -(6*R590? 5*R730) - 10* P R600–720)/Rmean Chlorite (USGS2007) Min690or700 [(R660? 3*R710)/4]/R690or700 Unknown Min480 [(R460? 2*R490)/3]/R480 Unknown
Fig. 2 a Selected reflectance spectra of the clastic sediment (prior to 1950) of Lake Silvaplana and b illustration for calculating variables. The figure shows measured reflectance data (black line), linearly interpolated reflectance continuum
between two maxima (blue line), points a and b used for calculating Relative Absorption Band Depth (RABD) and the area considered as Relative Absorption Band Area (red). BCA is the white standard based on barium sulfate
x3 (area = y1 9 (x3 - x1) plus the triangle between the points (x3, y1), (x1, y1), (x1, y3),
area ¼ðx3 x1Þ ðy3 y1Þ 2
We then have to subtract the area below the measured reflectance curve (black curve, grey bars Fig.2b). Because one measured value is covering a bandwidth of 10 nm, we have to subtract 10*R(max1?10; i.e.600) … 10*R(max2-10 i.e. 720) and
half of the bands of the first (max1) and the second (max2) maximum, respectively, i.e. 6*R(max1; i.e. 590)
and 5*R(max2; i.e. 730).
This results in the general formula:
RABA ¼ ðy1 x3 x1ð Þ þðx3 x1Þ ðy3 y1Þ 2
10XR max1þ10
ð Þ max210ð Þ 6Rðmax 1Þ
5Rðmax 2ÞÞ=Rmean
A comprehensive collection of reflectance spectra for different minerals can be found in the spectral library of the US Geological Survey (USGS 2007). Rein and Sirocko (2002) assigned the slope of the reflectance spectra of sediments between 570 and 630 nm to the amount of clastic material, i.e. the amount of illite or chlorite, in the sediment (Rein
2003). In Lake Silvaplana the spectra have a max-imum at 580–590 nm and a minmax-imum at 690–700 nm (Fig.2a). The maximum at 590 nm can be assigned to the amount of illite, biotite (both mica) and chlorite; the minimum at 690 nm can be attributed to the amount of chlorite (USGS 2007). The presence and relative abundance of these minerals in the sediments of Lake Silvaplana has been experimen-tally documented by XRD analysis of more than 400 individual varves (Trachsel et al. 2008).
To link the amount of different minerals and the mineral composition in the sediments of Lake Silvaplana to local climate variables, we followed the model of Trachsel et al. (2008). Three different geological formations with diagnostic bedrock com-position crop out in the catchment, which results in a different mineralogical composition of the sediment from different tributaries (Ohlendorf1998). Because the bedrock geology of the glaciated part in the catchment is relatively enriched in mica (biotite, illite) and chlorite, the sediment transported by glacial meltwater during warm summers carries that
signature. The relative proportion of these minerals can be detected by XRD or by specific characteristics of the reflectance spectra.
Prior to numerical analysis, all turbidites were removed from the sediment record, and the raw reflectance data were converted into a time series with homogenous (annual) intervals using re-sam-pling and linear interpolation procedures. To account for the minimum instrumental sampling resolution and for dating uncertainties, we applied 5 and 9-year running means to the time series, referred to as 5- or 9-year smoothed.
Numerical methods, calibration and climate reconstruction
In this section we describe the numerical methods applied in this study. First we tested the independence of the reflectance-derived variables and assessed the amount of redundancy in the multivariate data set. In a second step we assessed the potential and suitability of reflectance-derived variables for climate recon-structions and calibrated the variables against local meteorological data. We compared univariate and multivariate calibration approaches since it was not known a priori whether an individual reflectance variable or a combination of variables reveals better calibration statistics.
The technical resolution, i.e. aperture of the Spectrolino instrument sensor is 2.5 mm. It is not known, however, if the technical resolution is equal to the area measured (2-mm increments) on the sediment core. To evaluate the effective measurement resolution and assess the independence of the indi-vidual measurements we calculated autocorrelations of the reflectance-derived variables for the calibration period 1864–1950. The significance of the lag-1 autocorrelations was tested with a z-test (Venables and Ripley2002).
We then applied a Principal Component Analysis (PCA, Hotelling1933) to our data to detect similar-ities among the six reflectance-dependent variables (Table 1; two variables for illite, biotite and chlorite, two for chlorite and two ‘unknown’) and tested the significance of the PC-axes with the broken stick model (Frontier1976; Legendre and Legendre1998). Blass et al. (2007a) and Trachsel et al. (2008) have shown that the amount of clastic material (i.e. the mass accumulation rate, MAR) and the mineralogical
composition of the sediment reflect summer temper-ature and summer precipitation. Therefore, we used redundancy analysis (RDA, van der Wollenberg
1977) to assess the sensitivity of the reflectance spectrum-derived variables to all possible combina-tions of the three variables ‘MAR’, ‘temperature’ and ‘precipitation’ for 5- and 9-year smoothed data (Table3). The significance of RDA axes was assessed with permutation tests (Legendre and Legendre1998). MAR shows a low-frequency trend related to glacier length fluctuations and is mainly influenced by summer temperatures on shorter (inter-annual, decadal to multi-decadal) timescales (Blass et al. 2007a). Therefore, in all cases where RDA reveals an influence of MAR on the reflectance-derived variables, we detrended our data set to remove effects that are related to the low-frequency trend in the MAR. We used locally weighted regression (loess, Cleveland and Devlin 1988) to detrend our data. The span of the filter was set to represent 100 years.
To detect relations between reflectance-derived variables and climate variables individually, we cor-related unsmoothed, 5- and 9-year smoothed data with temperature and precipitation, using both raw data and loess detrended data. The degrees of freedom (DF) were corrected for autocorrelated time series according to Dawdy and Matalas (1964). Multiple testing was taken into account following Benjamini and Hochberg (1995). P values adjusted for autocorrelation are referred to as pc, and p values adjusted for
autocorre-lation and multiple testing are referred to as pca.
To produce a summer temperature reconstruction, we used inverse regression to calibrate those VIS-RS variables individually that are significantly (pca\
0.05) correlated with detrended instrumental summer temperatures. We then tested whether multiple linear regression (MLR) yields better calibration models than univariate calibration. Leave–one-out cross-validated RMSEP was used to choose the best model (e.g. Kamenik et al.2009). For the MLR model, we included an additional variable only if the RMSEP of the new model, with the additional variable, decreased by more than 5% (Birks1998). Because the focus of this study was on the reflectance spectroscopy method, we did not compare the performance of MLR to other multivariate calibration approaches.
To assess the performance of our 800-year reflectance spectra-based summer temperature reconstruction, we
compared our data with two independent reconstruc-tions based on documentary and early instrumental data (Casty et al.2005), and tree ring late-wood density data (Bu¨ntgen et al.2006).
Results
Reflectance-derived variables, PCA
The reflectance spectra (Fig.2a) of the sediments from Lake Silvaplana show distinct characteristics: a reflec-tance maximum at 580–590 nm is followed by decreasing reflectance with a minimum at 690–700 nm. The reflectance increases again towards 730 nm. A further local reflectance minimum is found at 480 nm. The six variables derived from the reflectance spectra and used in this study (Table 1) describe distinct characteristics of the spectra.
The time series of these six variables are shown in Fig.3a–f. Three of them (Ratio R590dR690, Ratio
R570dR630and min480; Fig. 3d–f) exhibit a long-term
trend, whereas no trend is detected for the other variables (Fig.3a–c).
The lag-1 autocorrelation of the spectrum-derived time series during the calibration period is significant (p \ 0.05) for all unsmoothed non-detrended data series as well as for unsmoothed detrended time series except for min480. After 5- and 9-year smoothing, the lag-1
autocorrelation is higher than 0.9 for all variables. When smoothing with filters higher than 9-year, smoothing no longer increases the lag-1 autocorrelation.
The results for the PCA are shown in Table2. For raw unsmoothed data, we found two significant PC-Axes, which explain 47 and 30% of the variance. The variables R590dR690, R570dR630,both indicative of
illite, biotite (mica) and chlorite, the trough590–730and
min690, both indicative of chlorite, have high positive
loadings in the first principal component PC1. PC2 loadings are positive for R590dR690 and R570dR630
(illite, biotite), but negative for trough590–730 and
min690(chlorite). Min690or700 and min480have
mod-erate negative loadings in PC1. Min690or700has a high
negative loading in PC2, whereas the loading of min480is moderately positive in PC2.
For the PCA of detrended spectrum-derived vari-ables we find one significant PC which explains 54% of the variance. The relationship of the variables R590dR690,R570dR630,the trough590–730and min690is
similar to the PCA for raw, non-detrended data (Table2).
Calibration and climate reconstruction
After describing the data by PCA, we used RDA to look for environmental parameters that might influ-ence the reflectance data. A series of RDA revealed highest influence of the summer season, temperature
(T) for months JJAS and precipitation (P) for months MJJAS.
For the raw reflectance data, T JJAS and MAR explain the highest proportion of variance, [25% for 5-year as well as for 9-year smoothed data (Table3). The proportion of variance explained by P MJJAS is considerably lower and amounts to 8.5% for 5-year smoothed variables and to 16.3% for 9-year smoothed variables. Combined T JJAS and MAR
Fig. 3 Raw VIS-RS data time series for variables a min690 or 700, b trough590–730, c min690, d R590dR690, e R570dR630and f min480.
explain more than 45% of the variance for 5-year as well as for 9-year smoothed data. Adding precipita-tion to T JJAS and MAR does not increase the proportion of explained variance. Thus, the third RDA axis is not significant. If only one explanatory variable is considered, i.e. excluding the effects of the other explanatory variables caused by multi-collin-earity, T JJAS showed highest values.
For the detrended VIS-RS data sets, the variance explained by T JJAS was higher, compared with
the raw data, whereas the variance explained by MAR and by P MJJAS was low. For the 5-year smoothed VIS-RS data, all of the three constraining data sets (temperature, precipitation and MAR) were significant (p \ 0.05). The same was found for the 9-year smoothed VIS-RS data. When combining explanatory variables, only the first RDA-axis is significant (p \ 0.05). The second and third RDA axes do not reach the significance level (p \ 0.05) when the explanatory variables are
Table 2 Loadings of the VIS-RS-derived variables on the significant principal component axes for non-detrended and non-detrended VIS-RS data
Significant correlations (p \ 0.05) between variables and axes are denoted with an asterisk
Non-detrended Detrended
Variable Correlation with axis 1 = Loading
Correlation with axis 2 = Loading
Variable Correlation with axis 1 = Loading
Min480 -0.125 0.38* Min480 0.19
R590dR690 0.88* 0.40* R590dR690 0.85*
Min690 0.84* -0.46* Min690 0.93*
Min690or700 -0.13 -0.92* Min690or700 0.29*
R570dR630 0.87* 0.42* R570dR630 0.86*
Trough590–730 0.77* -0.53* Trough590–730 0.90*
Table 3 Amount of variation (%) explained by T JJAS, P MJJAS or MAR and their combinations among six VIS-RS-derived variables
T JJAS MAR P MJJAS T JJAS ? MAR T JJAS ? P MJJAS P MJJAS ? MAR T MJJAS ? P MJJAS ? MAR Non-detrended data
5-year VIS-RS data 1 RDA 25.7%* 22%* 8.5%* 26%* (24.3%*) 28.3%* (23.7%*) 23.4%* (3%*) 28.8%* (24.1%*) 2 RDA (cum) 46%* (21.9%*) 32.2%* (6.9%*) 25.1%* (16.6%*) 49.1%* (2.8%*) 3 RDA (cum) 49.2% (17%*) 9-year VIS-RS data 1 RDA 33.4%* 28%* 16.3%* 33.8%* (33.8%*) 33.7%* (29.4%*) 29.1%* (16.4%*) 33.9%* (30.6%*)
2 RDA (cum) 61.6%* (28.25%*) 45.7% (12.29%*) 32.7% (16.4%*) 63.9%* (1.6%) 3 RDA (cum) 64% (17.6%*) Loess-detrended
5-year VIS-RS data 1 RDA 29.4%* 5.5%* 6.9%* 29.5%* (25.3%*) 35.4%* (28.6%*) 9.9%* (5.3%*) 35.4%* (26.2%*) 2 RDA (cum) 30.8% (1.4%) 35.5% (6%*) 10.8% (3.9%*) 37.5%* (6.2%*) 3 RDA (cum) 37.6% (1.4%) 9-year VIS-RS data 1 RDA 44.8%* 9.3%* 12.2%* 44.9%* (37%*) 46.8%* (36.6%*) 17%* (8.4%*) 46.9%* (31.2%*)
2 RDA (cum) 46.2% (1.4%) 46.9% (3.9%*) 17.6% (5.4%*) 48% (2.5%) 3 RDA (cum) 48.5% (1.2%)
The values in parentheses are the unique values of the variables (order according to the headline). Significant RDA-axes (p \ 0.05) and unique values of variables are denoted with an asterisk
combined. T JJAS again shows highest unique values.
Table4summarises the relationship between indi-vidual VIS-RS variables and climate variables. For raw (not detrended) reflectance data, the highest correla-tions are found between summer temperatures and the variables min690 and trough590–730 (unsmoothed,
5- and 9-years). While the correlation of trough590–730
with summer temperature fulfills the rigorous signif-icance level for 5-year smoothed data (pca\ 0.05
corrected for autocorrelation and multiple testing, marked with ** in Table4), all other correlations are significant for pc\ 0.05 (marked with * in Table4).
Reflectance variables are generally not signifi-cantly correlated (pc\ 0.05) with summer
precipita-tion, except 9-year smoothed R590dR690 and R570
dR630(significant at pc\ 0.05).
For detrended data, the highest correlations with summer temperature are found for R590dR690, R570
dR630 and trough590–730. These correlations are
sig-nificant (pca\ 0.05) for 5-year and 9-year smoothed
data series. For detrended precipitation, only one significant (pc\ 0.05) correlation was found after
9-year smoothing (R590dR690).
In summary, the variables trough590–730, and to
some extent R dR and R dR , have the
potential to serve as robust predictors for summer temperature. Although the variable trough590–730 is
also robust for raw (not detrended data), loess-detrending enhances the significance of the correla-tions with R590dR690and R570dR630. As a result of the
RDA and correlation analysis, a calibration for and reconstruction of ‘precipitation’ was rejected and no longer pursued.
Table 5shows the quality of the calibration (crite-rion: leave-one-out RMSEP) between detrended reflectance data and summer temperatures for univar-iate and multivarunivar-iate (2–6 variables) models. The multivariate calibration approaches yield consistently better results (smaller RMSEP) than models that are based on a single reflectance variable. Calibrations including two variables (trough590–730, min690or700)
resulted in a decrease of RMSEP by 15% compared to the best univariate calibration. Including three vari-ables in MLR (trough590–730, min690or700, min480)
resulted in a further decrease of RMSEP by 8%. When including more than three variables, the improvement of the calibration was always \5%. Thus the MLR model including three variables was found to be optimal and was used for the summer temperature reconstruction. The comparison between the instru-mental JJAS temperature data and the MLR calibration
Table 4 Correlations between T JJAS and P MJJAS and spectrum-derived variables for (a) unsmoothed, 5- and 9-year smoothed data and (b) loess-detrended un-smoothed, 5- and 9-year smoothed spectrum-derived variables
T JJAS P MJJAS
Un-smoothed 5-year 9-year Un-smoothed 5-year 9-year
Not detrended Min480 -0.34** -0.50 -0.58* 0.19 0.08 -0.04 R590dR690 0.23 0.60* 0.61 0.08 0.44 0.66* Min690 0.31* 0.70* 0.75* -0.07 0.12 0.08 Min690or700 0.05 0.00 0.07 -0.20 -0.25 -0.46 R570dR630 0.23 0.61* 0.60 0.09 0.39 0.56* Trough590–730 0.31* 0.73** 0.81* -0.01 0.16 0.13 Detrended data Min480 -0.31* -0.51** -0.63 0.13 -0.12 -0.36 R590dR690 0.26 0.72** 0.83** 0.04 0.38 0.50* Min690 0.26 0.66** 0.74 -0.07 0.21 0.26 Min690or700 0.06 -0.01 0.00 -0.18 -0.08 -0.17 R570dR630 0.24 0.71** 0.84** 0.05 0.35 0.41 Trough590–730 0.27 0.72** 0.82** -0.07 0.20 0.32
* p \ 0.05 taking into account autocorrelation
model is shown in Fig.4(apparent r2= 0.84; period 1864–1950: 9-year smoothed, detrended).
Figure5a and b show the reflectance-based sum-mer temperature reconstruction from Lake Silvaplana back to AD 1177 and the comparison with two independent regional summer temperature recon-structions based on documentary and early instru-mental data back to AD 1500 (Casty et al.2005) and tree-ring data back to AD 1177 (Bu¨ntgen et al.2006). The 200-year running correlations with the sum-mer (JJA) temperatures of Casty et al. (2005) were significantly positive (pc\ 0.1) back to 1715 (time
window 1615––1815) and remained positive for the time period back to 1500 (time window 1500–1700; Fig.5c). The comparison with the tree-ring-based (late-wood density) reconstruction of Bu¨ntgen et al. (2006) showed significant (pc\ 0.1) positive
corre-lations back to 1760 and between 1480 and 1550 (Fig.5d).
The decadal-scale variability of summer temper-atures is well captured. In particular, extremely cold
decades such as the decade AD 1810–1820, with strong negative volcanic forcing are obvious in the reconstruction. The running correlations reveal that, even for longer periods (e.g. AD 1550–1650), the reflectance-based reconstruction performs much bet-ter than the tree-ring reconstruction, when compared with the documentary data series. For other periods (e.g. AD 1200, 1450), there is a systematic offset between the lake sediment and the tree-ring time series. This offset is smaller than the dating uncer-tainties of the age-depth model of the sediment core, and therefore attributable to an imperfect chronology of the lake sediments.
Discussion
Reflectance-derived variables
Four of the six reflectance variables that were extracted from the reflectance spectra are indicative
Table 5 Comparison of leave-one-out cross-validated RMSEP (C) for detrended variables RMSEP
1 variable 2 variables 3 variables 4 variables 5 variables 6 variables R590dR630 0.142
R570dR630 0.136
Trough590–730 0.143
MLR Trough;
min690or700
Trough; min690or700;
min480
Trough; min690or700;
min480;R570d630
All except min690or700
All
0.116 0.102 0.100 0.098 0.098
First the RMSEP of the univariate calibration using inverse regression is shown for the three best variables (correlation with T JJAS pca\ 0.05). The RMSEP based on multiple linear regression models including two to six variables shows that including more than
three variables does not further improve the calibration model
Fig. 4 Multivariate calibration of detrended 9-year smoothed VIS-RS-derived variables (blue) with detrended 9-year smoothed TJJAS (red; r2= 0.84). The calibration model is based on multiple linear regression and includes three variables (Table5)
of the amount of chlorite and mica or combinations of these minerals in the lake sediment (USGS 2007; Rein 2003). A direct quantitative comparison of the variables with XRD measurements (Trachsel et al.
2008) was not possible for two reasons: (1) the XRD was measured in a semi-quantitative way that only allows the calculation of mineral ratios (peak inten-sity ratio); and (2) the reflectance-derived variables are influenced by the reflectance spectra of several minerals that overlie one another.
The high autocorrelation (lag-1) of five out of six time series is remarkable. This might be interpreted in three ways: (1) due to light scattering on the polyethylene foil and the wet sediment surface, the area integrated in the reflectance measurement might be larger than the aperture of the sensor (diameter = 2.5 mm) to which the light is emitted by the instrument; (2) the variables indicative of illite, biotite and chlorite and min690or700 (unknown)
have high persistence throughout the entire profile and reflect a multi-annual rather than an annual climate signal; (3) the autocorrelation is enhanced by the trend in the time series of the variables R570dR630,
R590dR690 and min690or700 (Venables and Ripley
2002). Most likely these three effects interfere with each other and it remains unknown whether the highest possible effective resolution of the VIS-RS data set is as small as the technical resolution (aperture = 2.5 mm) of the instrument.
The autocorrelation in the time series has consid-erable consequences for the subsequent statistical analyses (correlation and calibration). High autocor-relation decreases the number of independent obser-vations. As a consequence, the degrees of freedom DF need to be adjusted when significance levels are calculated (Trenberth1984). We recommend rigorous testing and careful assessment of significance levels rather than looking at correlation coefficients per se.
Fig. 5 a Comparison of the VIS-RS-inferred T JJAS recon-struction (multiple linear regression including 3 variables, blue) with the T JJA reconstruction based on early instrumental and documentary data (red; Casty et al.2005) and with b the late-wood density (mxd) based T JJAS reconstruction (green; Bu¨ntgen et al.2006). c 200-year running correlations between the VIS-RS-inferred T JJAS reconstruction and the T JJA of Casty et al. (2005) and d tree-ring based (late-wood density) TJJAS reconstruction (Bu¨ntgen et al. 2006). Significant
(pc\ 0.1) correlations are highlighted in black. For running correlation the year indicated is the central year of the 200-year time window. e 200-year running correlations between the T JJA reconstruction of Casty et al. (2005) and the TJJAS reconstruction of Bu¨ntgen et al. (2006). All data a and b are 9-year smoothed and LOESS-detrended (span 100 years), thus the reconstruction has skill in the decadal and multi-decadal frequency domains 9–100 year
PCA reveals large similarities among the first group of variables that describe the slope of the spectra (i.e. between R570dR630 and R590dR690
[Table2]). Although the segments are slightly dif-ferent from each other, they generally reflect the slope of the reflectance spectra between the global maximum, at about 580 nm, and the local minimum at about 690 nm. The variable R570dR630 is well
established in the literature and reflects the clastic compounds, in particular chlorite, illite or biotite (Rein and Sirocko2002; Rein2003). A second group consists of the variables min690 and trough590–730,
both indicative of chlorite (USGS 2007). These variables describe the absorption band depth and the absorption area of the major reflectance minimum at 690 nm. Thus, high similarity between the two variables is expected. The remaining two variables min480and min690or700(Table 1; indication unknown)
are different from the other two groups. These variables are poorly or not correlated with climate variables.
Calibration and climate reconstruction
RDA revealed a strong influence of MAR on the reflectance data series. MAR is subject to a low-frequency trend related to glacier length fluctuations (Blass et al. 2007a). To produce a summer temper-ature reconstruction that is not affected by the trends related to MAR, all variables (predictors and predict-ands) need to be detrended. RDA for detrended time series is clearly showing that temperature is the main factor influencing the reflectance-derived variables.
Significant correlations between summer temper-ature and reflectance-derived variables are found for min480, min690 and trough590–730 for raw (not
detr-ended) 9-year smoothed data and for R570dR630,
R590dR690 and trough590–730 for detrended 9-year
smoothed data (Table4). Besides the unknown variable min480, the other significant variables
con-sidered here are indicative of chlorite or mica (Rein
2003; USGS2007). It has been established that, due to the particular geological setting of the catchment (Spillmann1993), the amount of chlorite and mica is highest in the sediment transported by glacial melt-water (Ohlendorf1998; Trachsel et al.2008). During warm summers, glacier meltwater runoff is enhanced and more mica- and chlorite-bearing sediment from
the glaciated part of the catchment is transported into the lake.
Considering that this is, to our knowledge, the first methodological attempt to calibrate scanning reflec-tance (380–730 nm) VIS-RS data from minerogenic lake sediments with meteorological time series, the calibration statistics are remarkable. Individual reflectance variables correlate highly with instrumen-tal summer temperature data during the calibration period AD 1864–1950 (r [ 0.71 for 5-year smoothed series; r [ 0.82 for 9-year smoothed series; both highly significant pca\ 0.05, p corrected for
auto-correlation and multiple testing). The leave-one-out RMSEP of a multivariate calibration model, includ-ing the three variables (trough590–730, min690or700and
min480) is as small as 0.1C (r2= 0.84). However,
RMSEP is underestimated in autocorrelated time series when calculating leave-one-out RMSEP (e.g. Telford and Birks 2009). Such calibration statistics are rarely found in quantitative paleolimnology and show the great potential of the novel, rapid, non-destructive technique of scanning VIS-RS. It should be noted, however, that the quality of the calibration model and the calibration statistics, i.e. the final product used as a basis for the climate reconstruction, rely critically on an accurate chronology, in this case varve counts, on good knowledge of the mineralog-ical composition of the individual varves, here determined by XRD, and on a sound understanding of the multiple processes that lead to a specific climate signal in the varves, i.e. meteorology in the catchment, sediment transport and deposition.
The stable significant (pc\ 0.1) running
correla-tion with the summer temperatures of Casty et al. (2005) indicates the high stability of the calibration model back in time. The periods with systematic offsets on the order of a few years suggest that the correlation coefficients could be further enhanced by improving the chronology of the sediment core. In this study we left the chronology in its raw form. It is evident, however, that given the high temporal resolution of the data acquired, the accuracy of the chronology is critically important. We recommend that non-destructive scanning reflectance spectros-copy be performed on the same sediment half-core that will later be used to establish the age-depth model by varve counting of thin sections, radio-isotopes or other destructive methods.
Conclusions and prospects
We tested the potential of scanning reflectance spectroscopy in the visible spectrum (380–730 nm), measured directly (in situ) on sediment half-cores, to produce quantitative high-resolution (annual–sub-decadal) climate reconstructions from minerogenic lake sediments. For this methodological study, we chose annually laminated deposits from proglacial Lake Silvaplana because the characteristics of the sediments are very well known from previous studies using established analytical techniques. This knowl-edge allows a sound interpretation of the reflectance spectra measured.
Using MLR, we developed a calibration model that explained 84% of the variance of summer (JJAS) temperature during the calibration period 1864–1950. The correlation of individual variables (indicative of chlorite and mica) were as high as r = 0.84. Previous experimental work (XRD) has shown that the amount of chlorite and mica in individual varves is indicative of warm season temperature in Lake Silvaplana.
We used a series of numerical methods and statistical tests to extract a robust climate signal and proper calibration from the multivariate reflectance spectra data set. We first tested the independence of the individual measurements by calculating autocor-relation functions, and then assessed the degree of variance common to the reflectance-derived vari-ables, applying a Principal Component Analysis. We then tested the influence of several potential climate and sediment variables, i.e. temperature, precipitation and MAR, on the reflectance variables by means of Redundancy Analysis. In our case study of Lake Silvaplana we found that the MAR influences the reflectance data. Because MAR in Lake Silvaplana is subject to a low-frequency trend (glacier length variations), we also removed this trend from the reflectance-derived variables prior to calibration to retain a proper temperature signal.
We then tested whether one individual reflectance variable or a combination of variables would yield a better calibration with climate data. Thus, we com-pared the calibration statistics of systematic univar-iate and multivarunivar-iate calibration approaches and found that a calibration model using MLR including three variables performed best. We then applied this calibration model to reconstruct summer tempera-tures back to AD 1177. This reconstruction was
validated by comparing it to two independent regional reconstructions. Running correlation analy-sis showed high conanaly-sistency back to at least AD 1600 between all data series. The decadal-scale structure and amplitude of summer temperature variability is well captured. Shorter periods with cold summers, after large tropical volcanic eruptions, are evident, e.g. around AD 1256, 1275, 1450, 1600s and 1810–1820.
We propose the following recommendations for application of in situ scanning reflectance spectros-copy on mineroclastic sediments: (1) non-destructive reflectance spectroscopy should be performed on the same split core that is subsequently used to establish the age-depth model and chronology of the sediment core. This is most critical because high chronological accuracy is fundamental when calibrating the reflec-tance data set against climate data; (2) prior to interpreting the reflectance spectra, the general min-eralogical composition and geochemistry of the sediment should be measured by established analyt-ical methods (e.g. XRD). This provides the basis for a sound interpretation of the reflectance spectra with absorptions of the minerals that are actually present in the sediment.
This study shows the great potential for applying in situ scanning VIS-RS on mineroclastic sediments. From a methodological point of view, the robustness of reflectance characteristics should be further assessed. For instance, it remains unknown to what extent variables such as water content on the split core sediment surface, degree of oxidation of the sediment surface, or temperature influence reflec-tance measurements. Systematic tests are under way to assess the robustness, reproducibility and technical performance of scanning VIS-RS. However, one key advantage of this method is that it is inexpensive and data are acquired rapidly. Several thousand data points, which correspond to several meters of sedi-ment core, can be generated in one day. This enables systematic study of spatial variability among sedi-ment cores in a given lake (intra-lake variability), which is difficult to achieve with traditional analyt-ical methods. A further avenue of research could involve systematic testing of the method on a variety of lakes with different sediment types and proportions of clastic, biogenic and even chemical components. Although von Gunten et al. (2009) demonstrated the application of the new method on sediments from a
eutrophic lake, with high proportions of organic matter, the method still needs to be tested in a lake with volcanoclastic or chemical sediment (e.g. marl or evaporites), or in lakes with limestone bedrock.
Acknowledgments We thank Alex Blass, Thomas Kulbe, Michael Sturm and Alois Zwyssig for their help during fieldwork. We are grateful to Krystyna Saunders for her proof-reading of the English text. Mark Brenner and two anonymous reviewers made useful comments and suggestions, and helped improve the manuscript. Project grants were provided by European Union FP6 project ‘‘Millennium’’ (Contract 017008), NF-200021-106005/1 ‘‘ENLARGE’’ and the NCCR Climate.
References
Begert M, Schlegel T, Kirchhofer W (2005) Homogeneous temperature and precipitation series of Switzerland from 1864 to 2000. Int J Climatol 25:65–80
Benjamini Y, Hochberg Y (1995) Controlling the false dis-covery rate: a practical and powerful approach to multiple testing. J R Stat Soc Ser B (Methodological) 57:289–300 Bigler C, von Gunten L, Lotter AF, Hausmann S, Blass A, Ohlendorf C, Sturm M (2007) Quantifying human-induced eutrophication in Swiss mountain lakes since AD 1800 using diatoms. Holocene 17:1141–1154
Birks HJB (1998) Numerical tools in palaeolimnology–Progress, potentialities, and problems. J Paleolimnol 20: 307–332 Blass A, Grosjean M, Troxler A, Sturm M (2007a) How stable
are twentieth-century calibration models? A high-resolu-tion summer temperature reconstruchigh-resolu-tion for the eastern Swiss Alps back to AD 1580 derived from proglacial varved sediments. Holocene 17:51–63
Blass A, Bigler C, Grosjean M, Sturm M (2007b) Decadal-scale autumn temperature reconstruction back to AD 1580 inferred from the varved sediments of Lake Silvaplana (southeastern Swiss Alps). Quat Res 68:184–195 Bu¨ntgen U, Frank DC, Nievergelt D, Esper J (2006) Summer
temperature variations in the European Alps, A.D. 755–2004. J Clim 19:5606–5623
Casty C, Wanner H, Luterbacher J, Esper J, Boehm R (2005) Temperature and precipitation variability in the European Alps since AD 1500. Int J Climatol 25:1855–1880 Caviezel G (2007) Hochwasser und ihre Bewa¨ltigung anhand
des Beispiels Oberengadin 1750–1900. MSc thesis. Uni-versity of Bern, Bern
Cleveland WS, Devlin SJ (1988) Locally weighted regression– an approach to regression-analysis by local fitting. J Am Stat Assoc 83:596–610
Dawdy DR, Matalas NC (1964) Statistical and probability analysis of hydrologic data, part III: analysis of variance, covariance and time series. In: Chow VT Handbook of Applied Hydrology, a Compendium of Water-Resources Technology. McGraw-Hill, New York, pp 8.68–8.90 Francus P, Bradley RS, Abbott MB, Patridge W, Keimig F
(2002) Paleoclimate studies of minerogenic sediments using annually resolved textural parameters. Geophys Res Lett 29
Frontier S (1976) Etude de la de´croissance des valeurs propres dans une analyse en composantes principales: comparai-son avec le mode`le du baˆton brise´. J Exp Mar Biol Ecol 25:67–75
Heegaard E, Birks HJB, Telford RJ (2005) Relationships between calibrated ages and depth in stratigraphical sequences: an estimation procedure by mixed-effect regression. Holocene 15:612–618
Hotelling H (1933) Analysis of a complex of statistical vari-ables into principal components. J Educ Psychol 24(417–441):498–520
Kalugin I, Daryin A, Smolyaninova L, Andreev A, Diekmann B, Khlystov O (2007) 800-yr-long records of annual air temperature and precipitation over southern Siberia inferred from Teletskoye Lake sediments. Quat Res 67:400–410
Kamenik C, Van der Knaap WO, Van Leeuwen JFN, Goslar T (2009) Pollen/climate calibration based on a near-annual peat sequence from the Swiss Alps. J Quat Sci 24: 529–546 Kaufman DS, Schneider DP, McKay NP, Ammann CM, Bradley RS, Briffa KR, Miller GH, Otto-Bliesner BL, Overpeck JT, Vinther BM (2009) Recent warming reverses long-term arctic cooling. Science 325:1236–1239 Legendre P, Legendre L (1998) Numerical ecology. Elsevier,
Amsterdam
McKay NP, Kaufman DS, Michelutti N (2008) Biogenic silica concentration as a high-resolution, quantitative tempera-ture proxy at Hallet Lake, south-central Alaska. Geophys Res Lett 35
Michelutti N, Blais JM, Cumming BF, Paterson AM, Ruehland K, Wolfe AP, Smol JP (2009) Do spectrally inferred determinations of chlorophyll a reflect trends in lake tro-phic status? J Paleolim (online)
Ohlendorf C (1998) High Alpine Lake Sediments as Chronicles for Regional Glacier and Climate History in the Upper Engadine, Southeastern Switzerland. PhD thesis. ETH Zu¨rich, Zu¨rich
Rein B (2003) In situ Reflektionsspektroskopie und digitale Bildanalyse–Gewinnung hochauflo¨sender Pala¨oumweltda-ten mit fernerkundlichen Methoden. Habilitationsschrift. University of Mainz, Mainz
Rein B, Sirocko F (2002) In situ reflectance spectroscopy– analysing techniques for high-resolution pigment logging in sediment cores. Int J Earth Sci 91:950–954
Rein B, Lueckge A, Reinhardt L, Sirocko F, Wolf A, Dullo W (2005) El Nino variability off Peru during the last 20,000 years. Paleoceanography 20: PA 4003
Rosen P, Dodbakk E, Renberg I, Nilsson M, Hall R (2000) Near-infrared spectrometry (NIRS): a new tool for infer-ring past climatic changes from lake sediments. Holocene 10:161–166
SMA (2002) Jahresbericht der Meteo Schweiz. Annalen der Meteo Schweiz
Spillmann P (1993) Die Geologie des penninischen-ostalpinen Grenzbereichs im su¨dlichen Berninagebirge, PhD thesis. ETH Zu¨rich, Zu¨rich
Telford R, Birks HJB (2009) Evaluation of transfer functions in spatially structured environments. Quat Sci Rev 28: 1309–1316
Trachsel M, Eggenberger U, Grosjean M, Blass A, Sturm M (2008) Mineralogy-based quantitative precipitation,
temperature reconstructions from annually laminated lake sediments (Swiss Alps) since AD 1580. Geophys Res Lett 35:L13707
Trachsel M, Grosjean M, Larocque-Tobler I, Schwikowski M, Blass A, Sturm M (2010) Quantitative summer tempera-ture reconstruction derived from a combined biogenic Si and chironomid record from varved sediments of Lake Silvaplana (south-eastern Swiss Alps) back to AD 1177. Quat Sci Rev 29:2719–2730
Trenberth KE (1984) Some effects of finite sample size and persistence on meteorological statistics. Part I: Autocor-relations. Mon Weather Rev 112:2359–2368
USGS (2007) Digital Spectral Library, splib06a.http://speclab. cr.usgs.gov/spectral-lib.html
van der Wollenberg AL (1977) Redundancy analysis: an alternative for canonical correlation analysis. Psycho-metrika 42:207–219
Venables WN, Ripley BD (2002) Modern applied statistics with S. Springer, New York
von Gunten L, Grosjean M, Rein B, Urrutia R, Appleby P (2009) A quantitative high-resolution summer tempera-ture reconstruction based on sedimentary pigments from Laguna Aculeo, central Chile, back to AD 850. Holocene 19:873–881
Wolfe AP, Vinebrooke R, Michelutti N, Rivard B, Das B (2006) Experimental calibration of lake-sediment spectral reflectance to chlorophyll a concentrations: methodology and paleolimnological validation. J Paleolimnol 36:91–100
Zolitschka B, Mingram J, van der Gaast S, Jansen JHF, Nau-mann R (2001) Sediment logging techniques. In: Last WM, Smol JP (eds) Tracking environmental change using lake sediments physical and chemical techniques. Kluwer Academic Publishers, Dordrecht, pp 137–153
